GenAI, Risk Systems and Hard Questions
Trying to answer “what good looks like?” and a rant about GenAI’s capabilities.
This last week or two I’ve been doing something in my day job, you know the one that pays the bills so that I can waffle away on here all day, that I usually refuse to do.
Over the years I have regularly been asked questions like “what’s best architecture for X?” or “what’s the best HPC setup for Y?” or “what does good look like for Z?”. My normal answer to these questions is a flat out refusal to answer. The truth is, there is no good answer. There are just far too many variables. Any possible answer would be caveated to such a degree that it would be useless to the inquisitor.
But I was asked just such a question in a context where it was harder to say no. Because I’m being paid to answer it 😆. Oh you want to know the question? “What does a good financial risk system look like?”. Now it’s a fair enough question to be asking someone who has seem more than the average person will see in several lifetimes but somehow despite that, I not something I have ever been asked. Till now.
It not easy and it has just confirmed my belief that I have been correct to not answer such questions in the past. A question to which the expected response was an A4 diagram has so far evolved into many, many diagrams and a white paper that is around 3000 words so far. And not yet complete.
I hadn’t planned to write this post today. It’s a little off the cuff. Even so It’s still better than the alternative.
You see, its Eid al Adha tomorrow and I was running a little short on time. Recently I’ve used ChatGPT for some competitor analysis and while it was a little sycophantic (as is to be expected) I was actually quietly impressed. Mostly by the sentiment analysis it managed to produce. It dug through not only the websites but also blogs and social media posts to provide its comparisons, drawing out strengths and weaknesses that went beyond a simple comparison of features or products.
Then yesterday I came across prompt suggestion that makes ChatGPT spit it what it knows about you. Again, I was relatively impressed by the summary it produced and what it had gleaned about me without explicitly being told.
In a moment of time poor madness, I figured I’d cheat. I just told ChatGPT I’m short of time and to write a post for me. Nothing else. No chat history. And it produced this tripe:
“Every major cloud provider will let you spin up a VM in seconds. But try asking them:
Why did this instance take 3x longer to complete the same workload as yesterday?
Why did costs spike when my usage profile didn’t?
Which layer actually failed when my job silently stalled for 4 hours?
Silence. Or worse — dashboards that show everything’s green.
High Performance Computing isn't just about speed. It’s about predictability, visibility, and control.
Yet most infrastructure feels like a black box, even in 2025.
At HMx Labs, we’re working on a different approach — one where every second of compute time is observable, accountable, and explainable.
More on that soon.
#HPC #CloudComputing #Observability #Infrastructure #FinOps #NextGenHPC”
I mean its close. It did a good job of copying my style. No, it did a great job of copying my style. To the point I wonder if others can now just instruct it to do produce content like mine. It’s about HPC (without being told to be). It’s even a little edgy and if you didn’t know what better it seems rather plausible.
But its garbage.
Sure, I like to engage in a little hyperbole as much as the next HPC nerd, but never once have I seen a 3x variation in cloud VM performance. And I should know. We’ve run tens of thousands of VM benchmarks runs at this point. Yes, there’s a variation. Nothing like 3x. Don’t be daft.
And if any of the above were to happen and your dashboard really were still green maybe you should call me after all. Which brings me to my next point, I am just not that pushy! Maybe ChatGPT is trying to tell me something about the state of my business after its competitor analysis….
But I only half write this stuff for marketing. I mostly write it because I have about a billion ideas screaming inside my head and if I don’t get some of them out into the open my head will explode. My poor employees, colleagues and friends will only put up with so much and so the rest (and the rants) falls upon you dear reader.
So, yea, AI still sucks after all.
I guess I should be grateful really. Perhaps people working in HPC and financial risk systems will be the only ones left with a job 🙄
Anyway... I digress. Back to my original point. The next time you ask someone a question like that, especially if you’re not paying them to answer, please don’t be offended if they refuse. They’re doing you a favour.
And you can probably safely ignore anyone that does give you an answer that doesn’t resemble a Tolkien novel. Much like the answers that ChatGPT and Gemini gave me when I asked them to give me that diagram, it will be garbage too.
AI will take over our jobs my #£%@.
This one got a bit ranty and I haven’t even been listening to Ed Zitron this time!
Eid Mubarak to all those that are celebrating today.
